[Bug]: wheel size exceeds 200MB

[nopepper]

Your current environment

-

🐛 Describe the bug

Seems like the main branch of vLLM is oversized, which is crashing the docker build process :/

=> ERROR [build 15/15] RUN python3 check-wheel-size.py dist                                                                                                           0.3s
------
 > [build 15/15] RUN python3 check-wheel-size.py dist:
0.309 Wheel dist/vllm-0.5.2+cu124-cp38-abi3-linux_x86_64.whl is too large (204.7188196182251 MB) compare to the allowed size (200 MB).
0.309 vllm/_C.abi3.so: 282.28890228271484 MBs uncompressed.
0.309 vllm/_punica_C.abi3.so: 158.21475219726562 MBs uncompressed.
0.309 vllm/_moe_C.abi3.so: 13.828536987304688 MBs uncompressed.
0.309 vllm/config.py: 0.06867504119873047 MBs uncompressed.
0.309 vllm/worker/model_runner.py: 0.06034278869628906 MBs uncompressed.
0.309 vllm/core/scheduler.py: 0.05170249938964844 MBs uncompressed.
0.309 vllm/engine/llm_engine.py: 0.049073219299316406 MBs uncompressed.
0.309 vllm/lora/layers.py: 0.046690940856933594 MBs uncompressed.
0.309 vllm/model_executor/layers/sampler.py: 0.046222686767578125 MBs uncompressed.
0.309 vllm/spec_decode/spec_decode_worker.py: 0.040375709533691406 MBs uncompressed.
------

 3 warnings found (use --debug to expand):
 - FromAsCasing: 'as' and 'FROM' keywords' casing do not match (line 128)
 - FromAsCasing: 'as' and 'FROM' keywords' casing do not match (line 138)
 - LegacyKeyValueFormat: "ENV key=value" should be used instead of legacy "ENV key value" format (line 225)
Dockerfile:123
--------------------
 121 |     # check the size of the wheel, we cannot upload wheels larger than 100MB
 122 |     COPY .buildkite/check-wheel-size.py check-wheel-size.py
 123 | >>> RUN python3 check-wheel-size.py dist
 124 |
 125 |     #################### EXTENSION Build IMAGE ####################
--------------------
ERROR: failed to solve: process "/bin/sh -c python3 check-wheel-size.py dist" did not complete successfully: exit code: 

[youkaichao]

which commit?

[simon-mo]

I would assume this happens when you add more version to the cuda supported arch list.

[nopepper]

@youkaichao latest commit 89c1c6a
@simon-mo yes, this is using the example command in the docs:

DOCKER_BUILDKIT=1 docker build . --target vllm-openai --tag vllm/vllm-openai

[youkaichao]

the reason might be that your building environment is different from ci. the ci wheel size check passed.

[slin1237]

my environment is a pretty standard A10 and H100 Ubuntu20/22
None of them passed.
Perhaps we should check on the CI environment and really try this out?
Users do not have clear visibility into how CI is configured

[youkaichao]

if you build wheel for your private usage, you can just remove that wheel size check. we have the limit because we need to upload it to pypi .

[slin1237]

I'm not building it for private use. I only ran docker build. I don't think a normal docker build should fail because of that

[gargnipungarg]

I am building on A10.4
Same error
Error: building at STEP "RUN python3 check-wheel-size.py dist": while running runtime: exit status 1

Did someone figure out other than removing this check?

[gargnipungarg]

Thanks